SC515

Optics has long been used for communication, but the recent rise of photonic analog “accelerators” may give it a role in computing as well. This course focuses on the principles and recent innovations of optical machine learning systems, i.e., optical systems that perform machine learning. To start, this involves defining the problem, decomposing machine learning into mathematical operations (nonlinearity, matrix multiplication, convolution, etc.) and identifying bottlenecks in digital hardware and opportunities for photonic acceleration. We highlight the theoretical advantages of optical computing, including high data bandwidth, support for multiplexing (quadrature, mode, time, wavelength, polarization) and low-loss data movement, and define the important figures of merit for any system, such as power consumption, throughput density, latency and bit precision. Based on this theoretical framework, we analyze concrete examples proposed in the literature, including schemes based on programmable interferometers, microring weight banks, coherent detection, diffraction and Fourier optics. These systems give a good sense of the photonic devices required for a high-performance machine learning system, which, depending on the scheme, can include phase shifters, beamsplitters, ring resonators, modulators and photodetectors. Material- and device-level design choices are evaluated, which usually align with the choices made in optical communications, but sometimes not. For perspective, we also cover historical approaches to optical computing that never worked out (or were supplanted by CMOS electronics) and discuss why recent technological innovations in photonics have leveled the playing field and may make optical machine learning viable in the future.

This course will enable participants to:

• Define and quantify the benefits of using optics for computing
• Compare the advantages and drawbacks of free-space and on-chip architectures
• List and define the figures of merit for an optical machine learning system
• Describe factors that affect or limit the scalability of optical computing hardware
• List ways that computation hardware can leverage optical degrees of freedom, e.g., amplitude, phase, mode, wavelength, polarization
• Distinguish optical machine learning from analog electronic approaches
• Calculate the energy consumption of an optical computing system

Open to students, postdocs and professionals with a background in optics / photonics who are interested in researching or developing optical machine-learning systems, or simply keeping up with the field. Knowledge of machine learning, computer systems and integrated photonics is a plus but not required.

Ryan Hamerly has a background in theoretical and quantum physics, but recently he has focused on integrated photonic systems for computing. He co-led the optical AI hardware teams at the Quantum Photonics Group at MIT (Prof. Englund) and NTT Research. In the past, he worked on OPO coherent Ising machines, coherent feedback control and gravitational wave physics. Hamerly received his BS from the California Institute of Technology in 2010 and PhD from Stanford University in 2016.